This as-told-to essay is based on conversations with Sarthak Dhawan, a 21-year-old cofounder of Turbo AI, based in New York. His words have been edited for length and clarity.
There was one month this year where we accidentally spent around $30,000 on Claude Code tokens, which was not great.
I don’t think of that month as a total mistake, though I’ve learned from it. That was a heavy shipping month, and the spend reflected that. High token months usually mean we’re innovating or trying new things.
My cofounder, Rudy, and I launched our AI learning tool app in January 2024 while we were in college, and we dropped out to pursue our business full-time last year. The way I think of it, our bottleneck is shipping speed, not a few thousand dollars in tokens. Slowing down to manage our token spending would’ve cost us way more in momentum.
When we started coding the app, AI code-generation models were pretty terrible. I wrote the code myself from scratch. Now, coding has changed so much, and the actual task of engineering feels so different.
AI changed my job from writing code to reviewing it
More of my day-to-day is trying to plan and describe how things are working at a high level and reviewing AI-generated code than actually writing my own. It’s a lot of vibe-checking stuff as I go.
I can feel my coding skills atrophying, and I think every engineer who uses Claude Code knows it’s doing that to them. The more Claude takes over the task, the less engineers will actually know their codebase.
Twenty years ago, for example, a group of engineers working on a codebase knew the ins and outs of that code deeply. They understood every single architectural decision really well because they wrote it.
As AI writes your code, the code base shifts into this entity that you don’t understand. At the same time, we can’t avoid using it, because then we’re going to be way less productive. The difference in speed isn’t comparable at all.
Our AI costs grew, but we don’t have a set AI spending budget
Because AI coding became more useful over time, we never set a tight budget for our AI token spending. We keep a loose eye on it, but there’s no approval process for using tokens.
We have about 10 people on our team, and costs have crept up as more of us use AI for more work. We’re fine with it as long as it’s driving output. We average around $20,000 a month on AI tooling costs for software development.
For measuring token costs per developer, we just say use whatever you want, and we will continue to do that for now.
Our bill hit almost $30,000 after I left one Claude setting on, but it was an easy fix
In April, our AI token bill jumped to around $30,000. Part of the reason our bill was so high was because I was working in fast mode in Claude without realizing it.
The setting on Claude Code, called fast mode, makes it feel faster, but it also costs way more per token. When I turned fast mode on and left it on, that’s when my bill skyrocketed.
Now, we flip fast mode on when we’re pairing and latency matters; otherwise, we keep it turned off. Switching out of fast mode barely made a difference in the speed of output. Normal mode is plenty fast, and the quality’s the same, so it was an easy save that we didn’t feel that deeply.
When it comes to saving tokens, we don’t overthink it
To save tokens, we grab the easy wins — default to standard mode, lighter models for simple tasks, don’t dump whole codebases into context — but we’re not stressing over every dollar. We crossed $13 million in lifetime revenue this year.
We’re still in the mentality that if it makes us more productive, it’s probably worth it for the business in the long run.
Do you have a story to share about how you’re spending AI tokens? If so, please reach out to the reporter at [email protected].
Read the original article on Business Insider
The post My startup accidentally spent $30,000 on AI tokens in a month. It was worth it to move fast — but we found a simple fix. appeared first on Business Insider.
